Researchers at Nvidia GEAR lab, Carnegie Mellon University, and UC Berkeley developed an agent framework called ENPIRE that lets AI coding agents autonomously schedule and iterate robot training; the robots learned to cut zip ties and insert GPUs into narrow, thin slots on a motherboard.
JAXA researchers reported in Science Robotics that the palm-sized spherical lunar rover LEV-2 (nicknamed SORA-Q) operated autonomously for more than 100 minutes after landing near Shioli Crater during the January 2024 SLIM mission, traveled an estimated 24 meters, and returned multiple images.
How does test-time scaling affect robots? We found that larger models, more thinking, and more context significantly help some prompts but not others. Similar to large language models, we can also train a router to achieve a better performance/latency trade-off! Paper: https://jadee-dao.github.io/direct/
Chelsea Finn's team fine-tuned a VLA so that a single policy can control any robot in a robot team, enabling multi-robot collaboration without explicit communication (the system described in the quoted tweet is named CHORUS). Results show that this method matches or outperforms training separate models for each robot or training a unified centralized model, and it scales easily to large teams. Paper: https://arxiv.org/abs/2606.12352
Investor Shi Ma speaks with Stanford PhD student Huang Wenlong (advisor Li Feifei) about his PointWorld work and 3D world models. PointWorld uses 3D point flows to unify state and action representations, mapping different robots' joint actions into point cloud changes via URDF to address inconsistent actions across embodiments; training used the DROID and BEHAVIOR datasets, with annotations relying on FoundationStereo, VGGT, and CoTracker3.
Flow reversal steering can "steer" diffusion-based vision-language-action models (VLAs) through high-level actions, such as high-level actions inferred from vision-language models (VLMs). It also lets us run reinforcement learning (RL) in the diffusion noise space, with high-level reasoning guiding exploration: think about the task first, then practice! 👇
Diffusion (or flow) makes for excellent policies, but training them with RL is notoriously hard: BPTT is unstable, and RL over diffusion blows up the horizon. In our new paper, we show how to optimize a flow matching actor with a "weird little trick"—using the identity matrix to "approximate" the Jacobian of the flow denoising process. 👇
PKU EPIC Lab proposes KPGrasp, a dexterous grasp generation method that combines an all-Euclidean 3D hand-keypoint parameterization with a Transformer flow-matching model to learn grasp priors from large-scale data, without contact losses or contact-based test-time optimization.
Why it matters: The paper presents a new contact-loss-free grasp generation method and specific results on two benchmarks, letting readers compare it with existing methods and assess its approach.
Scaling reinforcement learning to long horizons remains a major challenge. Long-horizon Q-learning (LQL) prevents the accumulation of bootstrapping error by constraining differences in the value function over long horizons. Compared with one-step TD and n-step returns, LQL shows significant improvement! Paper: https://arxiv.org/abs/2605.05812
PKU EPIC Lab released LIMMT (Less Is More for Motion Tracking), which it describes as the first data-centric study for physics-based simulated humanoid motion tracking. The method defines motion data quality along three dimensions: physics feasibility, diversity, and complexity. Training on less than 3% of AMASS achieves better tracking performance than training on the full dataset, and the team also cleaned web-sourced mocap data. The paper has been accepted to ICML 2026; see arXiv:2606.06953.
Why it matters: The paper presents a three-dimensional data quality framework and results from training on a small amount of data, offering a methodological reference for researchers working on humanoid motion tracking.
Today (June 3), I will be speaking at the CVPR "Test-Time Scaling for Computer Vision" workshop (1:30 pm PT) on how we can use test-time compute to improve the generalization of robot policies, Room 506. Also, I am speaking right now (in 5 minutes) at the "Foundation Model Deployment" workshop in Ballroom 2!
Through reinforcement learning post-training, EXPO-FT brought π0.5 to a 30/30 success rate on all 8 manipulation tasks tested, using an average of only 19 minutes of reinforcement learning data per task. The method targets challenging manipulation tasks, including routing string lights and inserting a power connector to light the string, potting a billiard ball, and placing a flower into a wine bottle. It is released as an efficient, reliable, open-source VLA fine-tuning method; the paper and video are at https://pd-perry.github.io/expo-ft/.
Zhang Wanruo, a 2021 direct-entry PhD student at Peking University's Open Laboratory of Intelligent Robotics, has had the paper "MXPose: Multiplex interactive learning for multi-view 3D Human Pose Estimation" officially accepted by Pattern Recognition (PR), a journal in the pattern recognition field. Her research direction is multi-view 3D human pose estimation.
A UC San Diego team created microrobot swarms made of living algae and nanoparticles that can switch between gathering and dispersing using light color: blue light makes them assemble into specified shapes according to a mask, while red light makes them disperse. In simulated wound tests, AI analyzed the wound and laser-printed a matching mask; under blue light the robots assembled onto medical tape in the shape of the wound, and after red light irradiation, over 90% were released to the target area in under two minutes. The authors note that the system currently relies on light, is limited to surface applications, and still requires loading the nanoparticles with drugs and testing on real wounds and living tissue.
This article is an interview with Dr. Jianlan Luo about his new work LWD (Learning While Deploying), which enables general-purpose robot policies to continuously perform online reinforcement learning from their own experience during real-world deployment. LWD combines offline RL pretraining with online RL fine-tuning to unify offline and online optimization objectives; the value function is trained separately with a VLM and uses both positive and negative samples (roughly over 60% positive and over 30% negative), and deployment data flows back into the training pipeline in about 40 seconds on average.
The Human Interactive Driving team at Toyota Research Institute (TRI) presented a paper at the AutoUI conference introducing a data-driven, user-centered scenario development method for studying the application of drift-capable control to road safety.
Researchers at EPFL published a method in Science Robotics that lets multiple robots with different designs complete the same task from a single human demonstration. The method does not rely on AI; instead, it analyzes the physical properties of three-revolute-joint robotic arms, marks safe regions within each robot's range of motion, and classifies three-joint robots into six categories based on shared physical constraints, writing those constraints into each robot's program. The authors call this kinematic intelligence.
An EPFL research team published the Kinematic Intelligence framework in Science Robotics, enabling robots to safely transfer skills to robots with different configurations from a single human demonstration. The paper is available at http://dx.doi.org/10.1126/scirobotics.aea1995.
Genrobot has open-sourced the Phase 3 UMI dataset, containing 13K+ hours of dual-arm data that is ready to use and supports sub-1ms frame-level synchronization and high-fidelity trajectory reconstruction. The data was collected by 3K+ contributors in 10K+ real homes, covering diverse everyday manipulation skills. Download at https://huggingface.co/datasets/genrobot2025/10Kh-RealOmin-OpenData.
Why it matters: The data scale and real-home collection background help readers judge whether this kind of UMI data can be used to train their own manipulation tasks.
Sony AI's autonomous table tennis robot Ace competed under the rules of Japan's professional table tennis league, defeating multiple elite and professional players; the research was published in Nature (DOI: 10.1038/s41586-026-10338-5).
Why it matters: The article explains Ace's competitive record and three key technologies, and places it alongside the milestones of Deep Blue and AlphaGo, helping readers judge where physical sports competition currently stands.
Sanctuary AI demonstrated zero-shot in-hand manipulation in the real world with its proprietary hydraulic hand: a policy trained entirely in simulation continuously reoriented a lettered cube to a target pose ten times in a row without dropping it. The operation relied solely on the fingertips with no palm support, requiring simultaneous stable grasping and approach to the pose target. The original text states that the hand has a relatively high number of active degrees of freedom enabling finger abduction, and that hydraulic actuation and compact hydraulic valves offer a path toward industrial-grade humanoid dexterous manipulation.
Global views on robots are not uniform, and this is shaping adoption. 🌎 Where is trust highest? 🤖 What is holding people back? ⚙️ Who is ready, and who remains cautious? Attitudes are determining who acts first. Explore the data from our study: https://hxgn.biz/4rOBrpD
Teams including Berkeley AUTOLAB propose CaP-X, an open framework for systematically studying Code-as-Policy agents in robot manipulation, centered on the interactive environment CaP-Gym and CaP-Bench, which evaluates 12 frontier models.
Why it matters: The original text presents a systematic evaluation framework for Code-as-Policy agents and the finding that test-time scaling can mitigate abstraction dependence, offering reference value for research on coding-agent manipulation control.
A research team at New York University has open-sourced YOR (Your Own Robot), a mobile dual-arm manipulation robot, with a bill of materials cost below $10,000. The project homepage is www.yourownrobot.ai, and the paper is at arxiv.org/abs/2602.11150.
Why it matters: An open-source dual-arm mobile platform under $10,000 offers a reproducible hardware path and measured data, making it suitable for evaluating the cost and capability trade-offs of building a mobile manipulation system in-house.
Physical Intelligence has added an RL token output to its latest model, π-0.6, which is learned rapidly through reinforcement learning by a very small actor and critic, allowing the model to be fine-tuned for precise tasks in hours or even minutes without training the entire model.
Dexterity published an article introducing system evaluation and production measurement for the Foresight packing agent. On the planning side, it was tested on four real warehouse box-type distributions, with density of 0.65–0.77 and stability of 99.7–100%, and distribution-free order statistics were used to give 95% confidence intervals; the article also found that randomly shuffling the box sequence causes local variance to collapse by about 1.4 times, making evaluation overly optimistic.
Why it matters: The original article fully discloses the evaluation method for the packing agent, including distribution-free confidence intervals and the evaluation bias caused by random shuffling; the method can be transferred to similar systems.
Physical Intelligence proposes a new method called RL tokens (RLT), in which a vision-language-action model (VLA) outputs a special token to interface with a lightweight online reinforcement learning policy, improving the speed and success rate of fine manipulation phases with only a small amount of real data.
Why it matters: The original clearly explains the interface design between the VLA and lightweight online reinforcement learning and provides speedup data on four tasks, so readers can assess whether this type of method applies to their own fine manipulation tasks.
18,000+ people. Nine countries. One question: what will spark the next leap in robotics? 🤖💥 Our Robotics Generation study shows that attitudes toward robots differ around the world. What will determine how quickly robots enter our lives is people, not cost or technology. https://hxgn.biz/4ugEYQd
Physical Intelligence proposed a method called Multi-scale Embodied Memory (MEM) that provides vision-language-action models (VLAs) with both short-term and long-term memory, enabling the model to complete tasks lasting up to 15 minutes, such as cleaning an entire kitchen or making a grilled cheese sandwich from scratch.
Why it matters: The article identifies memory mechanisms as the bottleneck for long-horizon tasks and presents a dual-scale design that stores video for the short term and language for the long term, an approach that can be transferred to other work.
Ignacio Montoya, a US Air Force veteran and biomedical engineering doctoral student, reported years of self-study at the Society for Neuroscience annual meeting: transcutaneous spinal cord stimulation (tSCS) combined with adaptive task-specific training may reactivate dormant neural pathways 12 years after a clinically complete spinal cord injury. The training used Wandercraft's fully self-balancing, hands-free Atalante X exoskeleton, with robotic assistance gradually reduced; his gait consistency and muscle activation improved, and endurance, posture, and bladder and bowel sensory function also improved.
Major announcement! For whole-body humanoid robot teleoperation Project homepage: https://zhutengjie.github.io/CLOT.github.io/
Toyota Research Institute (TRI) reviewed eight years of polymer electrolyte research conducted by its AMDD project at MIT, producing about 40 papers and 7 doctoral dissertations and building a robotic experimentation platform that shortened tests that originally took about a month to completion within a day. The team also developed a multitask graph neural network to accelerate MD simulations and released the largest polymer simulation dataset, but the authors believe that AI can currently accelerate optimization within a class of materials, and that breakthrough polymers relying on new physical mechanisms still require human creativity.
We found an emergent property of VLAs such as π0/π0.5/π0.6: as pre-training scale increases, the model learns to align human videos with robot data. This gives us a simple way to use human videos. Once π0.5 has learned to control a robot, it can naturally learn from human videos.
Physical Intelligence research found that in generalization scenarios where only human demonstrations appear, fine-tuning π0.5 with human first-person videos mixed into robot data alone improved policy performance by about 2x, without using any explicit transfer learning mechanism.
Why it matters: Experiments show that merely mixing human videos into fine-tuning improves the policy by about 2x on scenarios with purely human demonstrations, and this capability emerges as pretraining robot data scale grows.
A team led by Nitin J. Sanket, a professor at Worcester Polytechnic Institute (WPI), has developed palm-sized flying robots that mimic bats' use of ultrasound, aimed at search-and-rescue missions. AI software filters noise from the ultrasound signals, letting the robots detect obstacles within a 2-meter radius. The team chose low-power ultrasonic sensors commonly used in automatic faucets, and to suppress interference from propeller noise, designed 3D-printed structures placed in front of the robot to modulate sound waves, similar to the function of a bat's nose, ears, and mouth tissues. The robot currently works; the next step is to increase its flight speed.
Berkeley AUTOLAB proposes the STITCH 2.0 suturing robot system, with seven improvements including improved EKF needle pose estimation, new thread untangling methods, and an automated 3D suture alignment algorithm.
The paper "Generative Visual Foresight Meets Task-Agnostic Pose Estimation in Robotic Table-top Manipulation," a collaboration between SUSTech CLEAR Lab, LimX Dynamics, and the University of Hong Kong, has been accepted to CoRL 2025.
Why it matters: This research presents a concrete technical approach and experimental figures for learning manipulation from human video, helping readers understand the feasibility of training robot manipulation with video data.
Field AI Research Institute (FAIRI) is collaborating with NVIDIA to study robotic embodied question answering (EQ&A) in dynamic environments. The team uses the NVIDIA Omniverse platform, Isaac Sim, Omniverse Replicator, and the USD Search API to build physically accurate 3D office scenes, simulate changes in the environment over time, and plans to open-source this long-term embodied question answering dataset.
A Kessler Foundation case study shows that Wandercraft's self-balancing Atalante X exoskeleton helped a 64-year-old patient with chronic left-sided hemiplegia from stroke complete four weeks of robot-assisted walking training, with gait speed rising from 0.09 m/s to 0.19 m/s, cadence increasing 27.8%, and no adverse events. Its powered dynamic ankle joint can replicate physiological gait mechanics, improving weight-bearing symmetry and propulsion. The authors stress that more research is still needed to verify the results, which were presented at the American Physical Therapy Association annual conference.
A RAL 2021 paper proposes Confidence-Rich Mutual Information (CRMI), which computes information gain based on continuous probability distributions of occupancy grid cells and a beam sensor causal model (SCM), avoiding the shortcomings of the traditional inverse sensor model (ISM). Simulations and experiments show that CRMI-driven robots explore unexplored and ambiguous regions more finely, with quadratic complexity close to traditional OGMI, and can run online.